Dzluck/qwen35-2b-mix-merged
The Dzluck/qwen35-2b-mix-merged model is a 2.3 billion parameter language model derived from a LoRA fine-tuning of Qwen/Qwen3.5-2B-Base. This experimental merged checkpoint focuses on instruction following and formatting-oriented tasks, particularly markdown table generation and bash rewriting. It is intended for testing merged-HF deployment flows and comparing base vs. fine-tuned behavior, serving as an intermediate release for further format-fix fine-tuning.
Loading preview...
Overview
This model, Dzluck/qwen35-2b-mix-merged, is an experimental 2.3 billion parameter merged checkpoint based on Qwen/Qwen3.5-2B-Base. It was created through a LoRA fine-tuning process aimed at improving instruction following and specific formatting-oriented behaviors. While it can perform basic inference, its reliability for strict structured output is still under development.
Key Capabilities & Status
- Base Model:
Qwen/Qwen3.5-2B-Base - Fine-tuning: LoRA/QLoRA-style workflow, focusing on mixed instruction-following and formatting.
- Markdown Table Generation: Generally usable.
- Bash Rewriting: Partially usable.
- Structured Output: Currently unstable for strict YAML and JSON generation, as well as strict field-format constraints.
Recommended Use Cases
This model is best suited for:
- Testing merged Hugging Face deployment workflows.
- Comparing the behavior of the base model versus the fine-tuned version.
- Continuing downstream evaluation and serving as an intermediate release for further fine-tuning efforts to improve format reliability.
Limitations
- Structured output alignment is incomplete, with YAML and JSON formatting prone to drifting from requested constraints.
- Not recommended for tasks requiring highly reliable structured output, such as strict YAML/JSON schema generation or "output only code/config" style prompting.